How IBM's data science team quickens users' AI projects

Plenty of IT shops have machine learning and AI projects underway, or they have plans to launch one. But many struggle to come up with their first use case, or to properly marshal the technical and human resources to ensure ROI. It is Seth Dobrin's job, as head of IBM's recently formed data science elite team, to help corporate users achieve those goals.

Dobrin, officially vice president and chief data officer for IBM Analytics, discussed his team's experiences and observations with their first engagements comprised of more than 30 customers.

What surprises you the most as you work with users on their first AI or machine learning projects?

Seth Dobrin: One thing is that most enterprises will hire a group of data scientists and tell them to go do a bunch of 'data science-y' stuff. They won't have a strategy or come up with a value proposition, and they don't know how to operationalize data science in the context of the enterprise.

Seth Dobrin

[By comparison,] a manufacturing company wouldn't buy a new piece of equipment that didn't have an ROI proposition. My fear is that, in two years, these companies will say, 'We have invested in this data science thing, and we haven't generated any value from it.' Part of my job is to help them create a value proposition and develop a strategy around it.

How do you begin this process?

Dobrin: We walk them through our methodology to identify a specific use case and then break it down into component models. Most use cases worth doing are not a single model, but are comprised of many. Then, together, we build out everything they need for those models in an agile manner. At the end of the engagement, which doesn't go past 90 days, we leave them with a working model. But, more importantly, we have taught them how do [data science-based projects] in the context of the enterprise.

Once you jointly agree on a use case, what tools does IBM bring to the table?

Dobrin: For up to 90 days, clients have free access to our products and services. We have a data science platform [Data Science Experience Local] that can be deployed on any cloud, including Amazon's or Google's. We have Watson Studio on the IBM Cloud. Depending on their use case, we teach them how to use those tools.

With more advanced clients with shorter engagements, we show them the value-add of our tools, such as the ability to deploy an API, as well as to automatically retrain their models without having to consume their data scientists' time with that. We also show them all the collaboration functions, so they can connect automatically to data, as well as carry out some data exploration without having to write code.

What tools must users have to work with you on AI projects, particularly if they are not True Blue shops with things like IBM Cloud?

Dobrin: They have to be trained to learn how to gain access to the platforms, which they can access for free for 90 days. But they still need to provide the hardware on which we install the software, and we will stand up the environment. Some users are seeing [our platform] adds significantly more value over theirs or over the off-the-shelf components they might buy.

So, you stay with them until they launch their first AI- or machine-learning-based offering?

We think the future of data science is in the vein of Python.

Dobrin: Yes, and we feel confident we can do this in 90 days. We have a good blueprint to make this happen as long as they commit a group to work with us. One of our requirements for clients is they must at least match our resources. On the ground, I typically assign three or four people to an engagement. The clients at the start add three or four of their people, but so far quickly outnumber us three or four to one, which we take as a sign that we are adding value.

Typically, what sort of use cases interest customers the most?

Dobrin: We have engaged with oil companies doing oil exploration, retailers looking to optimize the efforts of sales personnel and manufacturer looking to streamline their manufacturing processes. Generally, it helps them build a machine learning or decision optimization solution, so they can build a predictive or prescriptive model they can [add to] their processes. We don't do the physical implementation to their software -- just deliver an API. If they want someone else to do the implementation, we refer them to IBM's services organization.

Do you require user organizations to have some proficiency with AI and machine learning, and do you generally find there are enough such people?

Dobrin: We do prefer that [users' data science teams] have some ability with a programming language. We think the future of data science is in the vein of Python, so we really like people who know Python. Data science is not something you can just [quickly] teach people. It is more than a three-month process to teach people to do data science.